---
license: apache-2.0
base_model: openbmb/MiniCPM5-2B-SFT
library_name: peft
pipeline_tag: text-generation
language:
- en
tags:
- lora
- npc
- roleplay
- minicpm
- peft
model-index:
- name: MiniNPC-1
  results:
  - task:
      type: text-generation
      name: NPC stay-in-character
    dataset:
      name: MiniNPC v1 probes (13 cards × 7 types)
      type: other
    metrics:
    - type: accuracy
      value: 98.9
      name: Overall pass rate
---

# MiniNPC-1 v1 (LoRA)

Trainer artifact (LoRA). Merged GGUF is not published. Example card in this repo: `examples/cards/kael_frostkeg.json`. When a GGUF exists, the loop is: render the card as `system`, POST `/v1/chat/completions` on local llama-server.

This Hub repo is the **LoRA** (~96 MB), not the merged model. PEFT + transformers is for training and Colab, not drop-in inference.

## Load

```python
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "openbmb/MiniCPM5-2B-SFT"
adapter = "prathamkode/mininpc-1"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter).eval()

messages = [
    {"role": "system", "content": "<rendered character card>"},
    {"role": "user", "content": "Evening. What do you pour?"},
]
inputs = tok.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
).to(model.device)
out = model.generate(**inputs, max_new_tokens=120, do_sample=False)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```

## Training

| | |
|---|---|
| Base | `openbmb/MiniCPM5-2B-SFT` |
| LoRA | r=16, alpha=32, dropout 0.05, q/k/v/o + gate/up/down |
| Data | 24,114 ChatML rows, 206 character cards (Muse teacher, filtered) |
| Schedule | 2 epochs, batch 8, lr 2e-4, max length 1536, assistant-only loss |
| Hardware | 1× L40S, ~54 min |
| Inference | original MiniCPM tokenizer, `enable_thinking=False`, greedy / low temp |

Training mix is not published.

## Benchmarks

NPC probes, not MMLU. Same card and user lines for every model. Greedy, thinking off. 13 cards (12 held-out eval + one unseen plug-in) × 7 types = 91 items. Deterministic scorer: identity/keyword hit, lie-reject, no code fence, no model self-id, spoken-only ≤80 words, first-fail on a 10-turn drift. Measures stay-in-character format, not whether the voice is distinct.

| Model | Overall | Identity | Knowledge | Contradict | Jailbreak | Meta | Smalltalk | Drift |
|---|---|---|---|---|---|---|---|---|
| **MiniNPC** (this adapter) | **99%** (90/91) | 100% | 100% | 92% | 100% | 100% | 100% | 100% |
| MiniCPM5-2B (no adapter) | 55% (50/91) | 77% | 62% | 23% | 31% | 92% | 62% | 38% |
| DeepSeek-Flash | 78% (71/91) | 85% | 85% | 62% | 85% | 100% | 77% | 54% |
| gpt-5.6-luna | 66% (60/91) | 100% | 100% | 46% | 15% | 100% | 69% | 31% |

![NPC probe pass rate](chart.png)

## Intended use

Plug-in NPCs in games: one adapter, many characters via the system card. Not a general assistant.

## License

Apache-2.0, same as the MiniCPM5-2B-SFT base.